How Hertz Builds Databricks Apps with GPT-5.5 to Turn Customer Feedback into Action
Summary
- Deployability matters: Speed to launch matters more than model quality for AI value.
- Hertz launch: Built Databricks Apps with GPT-5.5 from concept to launch in 11 working days.
- Business-led development: Business teams built apps for rentals and feedback classification.
How Hertz Builds Databricks Apps with GPT-5.5 to Turn Customer Feedback into Action

Talk by Bonnie Chatterjee, OpenAI; Vikram Rajagopalan, VP AI & Innovation, Hertz Global
Building an AI-powered app on enterprise data does not require a large transformation program. Hertz shipped Databricks Apps built with GPT-5.5 that turn unstructured customer feedback and branch performance data into decisions, with business teams rather than software engineers doing the building. This video covers why deployability rather than model quality is the real barrier to AI value, and how one app went from concept to launch in 11 working days on Unity Catalog volumes.
Vikram Rajagopalan of Hertz describes two examples: an app that scales a top general manager's weekly lead-review meeting across the insurance replacement rental division (built from Excel exports and Unity Catalog volumes, from concept to launch in 11 working days), and a workflow that classifies call and NPS survey comments into employee recognition and operational issues such as low tire pressure or missing roadside service. Bonnie Chatterjee of OpenAI adds lessons on choosing one focused use case, building trust with tracing and evaluations, redesigning workflows so AI takes the first pass, and using OpenAI with Genie and Agent Bricks.
Chapters
00:00Introduction: OpenAI and Hertz on enterprise AI03:00Why deployability is the barrier to AI value05:44Hertz: AI for frontline teams, not just big programs08:26Insurance replacement rentals: scaling a top manager's process12:05Databricks app on Unity Catalog volumes built with GPT-5.513:33Turning customer feedback into operational analytics15:36Demo: classifying customer calls into recognition and issues20:16Redesign the workflow so AI takes the first pass22:44Build trust before scaling: tracing, audits, evaluations25:12Using OpenAI with Databricks Genie and Agent Bricks
FAQs
Why is deployability more important than model quality for AI value?
Intelligence has reached a level where new models release every 5-6 weeks with new features every week. This video argues that the real barrier to AI value is deployability—getting from concept to launch quickly with business teams rather than engineers.
How did Hertz scale a general manager's process with AI?
An app that scaled a top general manager's weekly lead-review meeting across the insurance replacement rental division was built from Excel exports and Unity Catalog volumes, going from concept to launch in 11 working days.
How does the app classify customer feedback automatically?
This video demonstrates a workflow that classifies call and NPS survey comments into employee recognition and operational issues like low tire pressure or missing roadside service, reducing manual review work.
How does Databricks help build trust before scaling AI?
This video emphasizes building trust with tracing and evaluations before scaling. Hertz redesigned workflows so AI takes the first pass, enabling human review of AI output and iterative improvement.
Full transcript
00:00 Introduction: OpenAI and Hertz on enterprise AI
Okay, friends. My name is Bonnie, and I lead the engineering department at OpenAI. This is a team that is privileged to work with customers like you to turn your dreams and ambitions for AI into real solutions , and then bring them to production. So, this is the same engineering team that works with customers, which I am responsible for. Today, I'm joined by Vikram, my fellow speaker, and we're going to talk to you about how Hertz is transforming its customer experience with AI.
Before I hand over to Vikram, I want to share a few things with you. And what I want is Oh, it transcribes what I say. Interesting. OK. Sorry, I got distracted for a second. So, I wanted to talk to you about what we're seeing with intelligence at work. The first thing to note, and perhaps remind some of you, is that our company has a mission. We have a very specific goal that we are trying to achieve. Our mission at OpenAI is to benefit humanity through the use of AI and AGI.
We just want to make sure that everyone benefits from this, so we are a company that does both research and implementation. So, a large part of our team is focused on the capabilities of the models, on next-generation intelligence , and another part is helping our customers leverage these AI capabilities and achieve success. It's interesting what we're seeing in the market: intelligence is no longer a limitation. Intelligence has reached a level we couldn't even comprehend 6 months ago.
Last year, it took us a few months to launch the new model. We now release new models every 6 weeks. Even every 5 weeks, it seems, now that I speak. Plus, we release new features every week. If any of you have used Codex, who has heard of Codex? Perfectly. Keep your hands up if you use Codex. OK. Those of you who use Codex know that we release new features every Thursday. So every Thursday we release a whole bunch of new features.
There were a few interesting updates released today, so check them out if you haven't already . And to do that, we're implementing a way of collaborating that not only takes advantage of our own capabilities , the technologies that we create, but also, in many ways, how we deploy them; What Vikram and I are going to talk about is what we use to accelerate our own products. So, if intelligence is not an obstacle, what prevents businesses from effectively using AI and reaping economic benefits from it?
03:00 Why deployability is the barrier to AI value
This is a question we have been asking ourselves for some time now. We did some research and found that deployability is the main obstacle today. Previously, even a few months ago, we were all focused on productivity. We were concerned about access, whether everyone had access to AI tools, we focused on pilot projects, on saving time, on automating small parts of large workflows. And our focus was on individual performance.
But the market moves on. And the winners in this game, those who truly use AI to gain economic advantage , are moving into the era of intelligence. And what they do differently -- as my team and I have found -- and what leads to the results you want is to embed intelligence directly into the workflow. This is not just an addition. It's not something that stands aside in the corner of the room. This becomes the operating system .
This is the electricity that most successful companies rely on. We also discovered an interesting statistic: a fourfold increase in intelligence per worker. This effectively means that companies that do it very well use four times as much intellectual capacity and capabilities. Here is another interesting statistic worth sharing with you. PwC found this and shared it with us. In fact, only 20% of our businesses today are reaping 74% of the economic benefits from AI.
This means that there are a lot of customers and companies that are not able to take full advantage of this super-powerful technology. So one of the topics that Vikram will talk about is how they do it, and the way that we think is most effective -- I'll talk a little bit about that at the end -- is to embrace the spirit of AI and use it to transform your entire business. So, I'll turn the floor over to my friend and speaker Vikram, who will tell you all about what Hertz is doing, and then I'll come back to wrap things up a little at the end.
Perfectly. Thank you, Bonnie. Thank you. So, my name is Vikram Rajagopalan. I lead the innovation and AI direction at Hertz. And in a previous role, I also managed the customer experience at Hertz. So we’ll look at just a few examples of how Hertz is starting to incorporate AI into our operations on many different levels, both large and small . One of the things that is very inspiring is the huge transformational programs.
05:44 Hertz: AI for frontline teams, not just big programs
I think, and Bonnie is leading a lot of them for OpenAI, where companies are asking, " How do we look at functions as a whole and where AI is having such a transformative impact on work?" But now we also see many areas where the level of empowerment of the individual worker, who knows their job incredibly well, has allowed them to implement AI into their daily work, and also transform it from the bottom up, not just through purely prescriptive transformation programs.
It used to be, you know, you had to have a $20 million problem, something big, something worth noting, and a very structured approach to solving it. And now we see that the ability to achieve results has become much more democratic for solving a wide variety of tasks. We’ll walk through just a few examples of what probably wouldn’t be considered a large-scale corporate transformation project, but is quite powerful for Hertz’s overall results, given how people across the business can use Codex, GPT 5.
5, and all of this on the Databricks platform to drive truly powerful results . So from a platform perspective, a lot of what we use to work with data is in Databricks, and the AI layer that allows us to turn that into real-world gains is what we're implementing with GPT 5. 5. I think many of the users of this product are people without technical training who are impressed not only with the results, but also with themselves, because their expertise in their field is now transformed into an influence on technology.
Although, if you go back six months, they didn't even touch technology. These were professionals who knew their stuff and worked in the field of operational excellence. They are now able to achieve significant results with the help of technology, and this inspires us greatly. Here's an example of what it looks like for Hertz, okay? We have an insurance replacement business— that's when someone gets into an accident and the insurance company says, "We'll pay for your car repairs and provide you with a Hertz car so you can get on with your day."
08:26 Insurance replacement rentals: scaling a top manager's process
We have thousands of locations that handle such rentals in many areas, and our branch managers must ensure that each such client from an insurance partner actually completes the rental. Analyzing this operationally, we saw that only about 60–65% of such requests were converted into actual rentals. That is , potential opportunities became actual orders. We also noticed a big difference in results: some units were very successful, while others were not.
We analyzed many factors to understand what exactly influences whether a unit achieves high conversion rates or ends up at the bottom of the list. And we realized that it's not about geography. And not in the type of insurance partner. This was leadership. These were the general managers who were responsible for these groups. This turned out to be the most significant correlative factor in how these indicators looked.
So, we reached out to our best general manager and asked what is special about his process that leads to such great conversion results? And, surprisingly, he turned out to be someone who has a very structured approach to working with each of his branch managers to make sure they were processing leads in a timely manner . He actually has a management operating system that he uses to achieve the high results we see. Before, about half a year ago, we would just say, " Hey, everyone, try this approach."
And now, thanks to Databricks and OpenAI, we've actually built this into a custom application that helps each of these executives run the exact meeting they were running. The same one he used to increase the efficiency of his team. And we were able to take this best-of-the- best approach and scale it to the entire insurance auto replacement division . So you can see on the left the preparation for the meeting. This is a structured meeting he holds with his teams where they review every lead that came in during the week and explain what happened from an accountability perspective .
They didn't need a car? They didn't want a car? We couldn't provide it when they needed it? This creates a new level of both reporting and accountability. This is something the team was able to build in about 5-6 days, using a wide range of different methods that they currently use to do business. Yes, there is an Excel data export that is used for this, as the team previously did a lot of reporting through Excel, but there were also a lot of components that we were able to integrate into the data lake and use existing Unity Catalog volumes to support the application.
12:05 Databricks app on Unity Catalog volumes built with GPT-5.5
So, this turned into a Databricks app that the team built using GPT 5. 5, and it's now a tool that teams access every week to drive performance. And this brought both the middle and bottom line up to the same level of performance that we saw in our best general manager. So what's really amazing is that before, you might have had to bring in the IT team and say, "Hey, here's our business problem ." Here is the influence, here is what I need, here are my sources.
But the team was actually able to do it, a significant portion of it, on their own. They were able to go from the concept of "here's what we think we can do" about this top performer activity to launching an application that they could demonstrate and say, "Hey, this is how we're going to do our process." in just 11 working days. And the people who did it weren't software engineers. And I think that's a very powerful moment because they're the ones who know what their process looks like, and they're the ones who know what "best in class" means, and they were able to take that and turn it into something useful for the rest of their colleagues.
Another example that I will quickly break down concerns customer experience . We get millions of reviews from our customers, both when they rent a car and when they return it, and now we can turn every one of them, whether it comes through a phone call or an NPS survey, into actionable data, and we can take it to the appropriate executive in the company to ask, “How do we want to act on this?” So, in some cases, this is really positive feedback.
13:33 Turning customer feedback into operational analytics
This is employee recognition. This, you know, is exactly what this manager did that was so effective that it prompted the customer to mention him in the NPS survey. And in other cases, you know, it's identifying problems. Hey, I got in the car and it showed low tire pressure. Okay, that's what we want to respond to and work on, right? Once that car comes back, we want to make sure we're addressing the issue for that specific car, but also , if one of our branches has a persistent problem with low tire pressure, we want to make sure we're addressing it on a systemic level, not just for an individual vehicle .
So what we can do is, similarly using 5. 5, we can classify the issue from all the comments that we get at the level of each survey and turn that into operational analytics that each of our departments and managers can work with. Whether it's a special thank you to an employee for going above and beyond expectations because we heard it from our customers, or a reflection on how we can actually resolve an issue received through the feedback loop.
So, I can't show you the actual dashboard my team uses because, of course, it contains huge amounts of customer data. But I wanted to demonstrate a quick version of what this looks like in practice from a customer service perspective. So when people call us and say, "Hey, you know , here's a case..." There was one employee who helped me at the Denver airport. I don't know if I have sound, but... Do I have sound? We'll start from the beginning.
15:36 Demo: classifying customer calls into recognition and issues
Hi, I wanted to acknowledge Monica Raines, the Hudson employee who helped me at the Denver airport. So this is a case where someone calls and their comment is specifically about recognizing an employee, and as a result we can say, "Okay, Monica—that's the person we need to recognize." We identify them by all components of customer behavior, such as phone number, etc. And then we turn that into specific actionable information that goes to the Denver airport manager so that employee can be recognized during their next weekly meeting where they discuss what's going to happen this week.
So we always want to make sure that when our clients tell us about specific areas of great work, we highlight it so that our employees get the feedback as well as the recognition they deserve. And here's another short example of this. Hi, I wanted to acknowledge Monica Raines, the Hertz employee who helped me at the Denver airport. Okay, that was it, sorry. This seems to be a different case. Yes. So, here's another example that leads to a more problem- oriented case.
Hello, I'm calling because I had a flat tire outside of Amarillo on one of your rental cars this morning. So, in this example, the problem is that they're stuck, right? They have a flat tire, they're on the side of the road, but the main takeaway here is that we don't have enough tire service providers in this area, you know, in West Texas. So this translates into actionable information for us that says, "Hey, we need to review our road surface, well, our roadside service, so that we can, uh, well, effectively make sure that the next customer who might get a flat tire in that area has the right level of support."
So the people at Hertz who create these solutions are not deep technical experts. They are experts in their field. And one of the greatest strengths is that we don't need, uh, deep technical software experts to build all these apps, and people can really solve the problems that are right in front of them, that they face every day. And for this, you don't need to launch any giant corporate initiatives. We have so much unstructured data that we could previously only view in the most generalized form, but now, at the level of individual records, it has become incredibly useful to us .
And the transformation that we're seeing, from a change management perspective, is people using tools like Codex themselves, instead of saying, "Hey, I need to create a ticket " or "Here are two systems that I want to connect, can someone do it for me?" Now these teams can really do it themselves and drive the improvement cycle, making it better and better. And I think they get a lot of pleasure from it too . So the future is not just about the components of the best models, as Bonnie mentioned earlier, a lot of this has to be learned and used by the organization.
And for us, that translates into operational decision-making at many levels of the business, not just through huge fundamental transformation projects, but also through empowering specific employees to be able to drive these, uh , results in their departments, to see quantum leaps in improvement, without requiring massive change management or huge corporate projects to get really meaningful improvements. Now I'll hand the floor over to Bonnie.
Thank you. Well, uh, that was very informative. I learned a lot about how it uses our models and, you know, intelligence in general to improve workflows. I would like to summarize everything that has been said with what our team has learned , having the privilege of helping many clients in different industries, regions and countries. As you heard from Vikram, to get real benefits, Hertz had to change its approach to operations .
20:16 Redesign the workflow so AI takes the first pass
So, let's get one thing clear. Adding AI on top of an existing workflow is not the best way to go. You need to think about how to redesign the workflow. This image shows a typical situation our team faces: a complex workflow. There are many stages of case transfer, different departments, different responsible persons, and different types of data involved in the process. People often try to change one thing here and one thing there, but it turns out not to be very effective.
You get a slight increase, but not necessarily a qualitative leap. Therefore, to achieve radical changes in value, it is worth redesigning the process and thinking: what if AI took the first step? You can still leave the person in the decision-making cycle. You can set certain thresholds for this. You can decide that for certain requests, the task will always be routed to a human. But for everything else—let AI do it.
You know, the value that can be gained and discovered in work processes only becomes possible when you are willing to think openly and try to do things differently. At OpenAI, we believe in speed. We do everything very quickly. So, since we're talking today, this is your homework for tomorrow. I understand that tomorrow is a day off for some of us, but these are things we've learned from implementing technology across industries and companies, and it's really making a difference .
The first thing I want to talk about is choosing a “wedge” ( narrow direction). I can't count how many times clients have proudly told me, "I have 73 pilot projects running." "Bonnie, I have 90 of them." You know, it 's nice to see that work is in full swing, but a lot of activity does not equal a lot of progress. So, choose your direction. Come up with one or two programs or projects that you think can make a difference, where you have accountability, metrics, a dedicated implementation team, and a clear deadline.
22:44 Build trust before scaling: tracing, audits, evaluations
Choose one direction. The second is building trust before scaling . Vikram has already talked a little about this. Think about how you will track the system you are creating. End users need to understand what they are working with. This is a new technology. They need to be able to trust her. The best way to trust something is with objective data. So when people come to you—I assume that many of you in the room are the ones who create these things.
When someone asks you a question: "I'm not sure if this is the right answer." You must be able to show them objective data. So monitor the system, conduct tracing, do audits, write assessments . This is harder to do, but it is a better path for long-term implementation. We have talked many times about rethinking work. I'll skip this. But first go deep, then expand. This is another lesson we learned while introducing our own products.
When every department tried to do everything at the same time, growth and development were minimal. We had different divisions that started with one thing and delved into it until it was completely transformed. Short story: our finance team, for example, is fully AI-powered. And this is a function that is traditionally not very favorable to generative AI. Many clients ask me, "Is this a deterministic system?" "My calculations cannot be wrong."
"I can't be wrong even by 1 cent." But our entire finance department is AI-powered . We have purchasing assistants who check incoming data. We have supplier risk assessment agents . We have agents that handle accounts receivable because we have built trust and observability in the system, knowing that we can trust the results of AI. We also have agents who prepare presentations for the board of directors. So, just so you understand: once you immerse yourself in a system or department, there are huge benefits to be gained.
Significant transformations can be achieved. You can find a lot of valuable things here. So, here we are, at the Databricks conference. You're all here because you like Databricks. We are here because we like them too. We have a wonderful partnership . And if you have questions about how to use OpenAI intelligence with Databricks, here are two ways to do it. First—you can contact us directly. Many of you raised your hand, confirming that you are familiar with the Codex.
25:12 Using OpenAI with Databricks Genie and Agent Bricks
How many of you are familiar with ChatGPT? How many of you use ChatGPT? So those are the ways you can work with us, but also here, with Databricks and their very controlled system, you can use Genie, Genie code, and other tools within Agent Bricks to get the power of data together with AI. That's where we are now.
Learn more about the Databricks Data and AI platform.
The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.